Croatian speech-to-text API: 11% word error, measured

Measured accuracy

Model	Word error rate	Character error rate	Clips with no errors	Empty transcripts	Speed
Whisper Large v3 `openai/whisper-large-v3`	11%	2.6%	5 of 30	0 of 30	0.068× real time
Parakeet TDT 0.6B v3 `nvidia/parakeet-tdt-0.6b-v3`	10%	3.4%	7 of 30	0 of 30	0.056× real time

Good. Usable with light review. Expect to correct a word here and there, mostly names and numbers. The clips are clean, single-speaker read speech, so treat this as a best case for your own recordings. How we measured, and all 83 languages.

What a transcript looks like

The clip with the median error rate from the test, reference text against what the API returned.

Spoken	Toplina špilja nije tako ujednačena kao kod velikih špilja na Zemlji koje često održavaju poprilično stalnu temperaturu, već je dosljedna s obzirom na to da su to veliki otvori u tlu”, kaže Glen Cushing iz astrogeološkog tima Američkog geološkog zavoda (USGS) i Sveučilišta Sjeverna Arizona koje se nalazi u Flagstaffu u Arizoni.
Whisper Large v3	"Toplina špilja nije tako ujednačena kao kod velikih špilja na Zemlji koje često održavaju poprilično stalnu temperaturu, već je dosjetna s obzirom na to da su velike otvore u tlu", kaže Glenn Cushing iz astrogeološkog tima Američkog geološkog zavoda USGS i Sveučilišta Sjeverna Arizona koje se nalazi u Flagstaffu u Arizoni.
Parakeet (its median clip)	Spoken: „On [Wales] nam je načelno lagao od samog početka. Prvo se pravio da za to postoje pravni razlozi. A zatim se pravio da nas sluša, sve do samog brisanja fotografija.“ On, Wales, nam je načelno ulagao od samog početka. Prvo se pravio da zato postoje pravni razlozi, a zatim se pravio da nas sluša sve do samog brisanja fotografija.

Language detection

Send language=hr. Of 10 clips sent without it, only 8 were identified as Croatian; the others were labelled English (2). When detection is wrong, Whisper writes the transcript in the language it guessed.

Which model for Croatian

Parakeet TDT 0.6B v3 was the more accurate model on these clips (10% against 11%) and it is also the faster one. Start with nvidia/parakeet-tdt-0.6b-v3. Side by side.

Quickstart

The request is the OpenAI transcription request with a different base URL. language is optional in the API but, as shown above, not optional in practice for Croatian.

cURL

curl https://api.inferenceapis.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $INFERENCE_API_KEY" \
  -F "model=nvidia/parakeet-tdt-0.6b-v3" \
  -F "language=hr" \
  -F "file=@recording.mp3" \
  -F "response_format=verbose_json"

Python · openai SDK

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.inferenceapis.com/v1", api_key=os.environ["INFERENCE_API_KEY"])
with open("recording.mp3", "rb") as f:
    r = client.audio.transcriptions.create(model="nvidia/parakeet-tdt-0.6b-v3", file=f, language="hr", response_format="verbose_json")
print(r.text)

Subtitles and timestamps

response_format=srt or vtt returns a subtitle file with cues of up to two lines, broken at sentence ends and pauses. verbose_json with timestamp_granularities[]=word returns a start and end time for every word if you want to build your own.

Cost

$0.002 per minute of audio on either model: $0.12 for an hour, $12 for a hundred hours. Duration is measured from the file you upload. There is no daily cap on minutes and no subscription; enough to run your own Croatian test.

FAQ

How accurate is Whisper for Croatian?

On 30 clean read-speech clips it made 11% word errors, about one word in 9. Usable with light review. Expect to correct a word here and there, mostly names and numbers. Real recordings with noise, accents or several speakers will score worse.

Does language auto-detection work for Croatian?

Not reliably. Of 10 clips sent without a language, 8 were identified as Croatian; the rest were labelled English (2). A wrong label means the transcript is written in the wrong language, so always send language=hr.

Whisper or Parakeet for Croatian?

On the same clips Whisper Large v3 scored 11% and Parakeet 10%. Parakeet was more accurate here and is also the faster model, so start with it. Both cost $0.002 per minute.

Can it translate Croatian speech to English?

The API returns Croatian text. For English, transcribe first and translate the text with a chat model on the same key.

Related languages: Serbian (21%) · Bosnian (13%) · Slovenian (24%)